yinanhe / action-recognition

Archive action recognition by PyTorch depend on CNN and LSTM

Action Recognition

The network are based on CNNs and LSTMs. A little change.

简体中文

Environment

Python 3.6.5 (Anaconda 5.2.0):

https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/

pytorch 0.4.1 (CUDA 9.0):

https://pytorch.org/

Data

Use the following command to download the example data:

python data/download.py

Or download directly from Google Drive and unzip.

The training dataset structure:

~/
  data/
    train_data/
      .../ (directories of class names)
        .../ (directories of video names)
          ... (jpg files)
    valid_data/
      .../ (directories of class names)
        .../ (directories of video names)
          ... (jpg files)
    save_model/
      save_100.pth

The test dataset structure:

~/
  test_data/
    .../ (directories of class names)
      .../ (directories of video names)
        ... (jpg files)

The data structure of the video training dataset:

~/
  video_data/
      .../ (directories of class names)
        ... (avi files)

Use video2jpg.py under the utils folder to convert avi to jpg:

python utils/video2jpg.py video_root_directory image_root_directory

Use split_data.py to move some of the training dataset to a validation dataset:

python utils/split_data.py train_root_directory valid_root_directory

Train

Use the following command to train the model:

python train.py train_data_directory

For default:

python train.py data

Optional parameters of command:

--model DIR           path to model
--arch ARCH           model architecture (default: alexnet)
--lstm-layers LSTM    number of lstm layers (default: 1)
--hidden-size HIDDEN  output size of LSTM hidden layers (default: 512)
--fc-size FC_SIZE     size of fully connected layer before LSTM (default:
                      1024)
--epochs N            manual epoch number (default: 150)
--lr LR               initial learning rate (default: 0.01)
--optim OPTIM         optimizer (default: sgd)
--momentum M          momentum (default: 0.9)
--lr-step LR_STEP     learning rate decay frequency (default: 50)
--batch-size N        mini-batch size (default: 1)
--weight-decay W      weight decay (default: 1e-4)
--workers N           number of data loading workers (default: 8)

Test

Use the following commands to test:

python test.py model_directory test_data_directory

For default:

python test.py data/save_model/model_best.pth.tar data/valid_data

Predict

Use the following commands to predict:

python predict.py model_directory predict_data_directory

For default:

python test.py data/save_model/model_best.pth.tar data/predict_data

Network

This project is based on CNNs and LSTMs.